docs(decisions): D24 — autonomy green-light (lift M2 gate for build-out + delegated controller-merge) #110

Merged

pdurlej merged 1 commit from claude/decisions-d24-autonomy-greenlight into main

2026-06-23 08:01:28 +02:00

claude commented

2026-06-23 07:42:08 +02:00

Collaborator

What

Records your 2026-06-23 A/B/C decisions as D24, so the bold autonomy push has authoritative, auditable bounds (and codex knows exactly how far "go big" goes).

The decision (two parts)

M2 gate lifted for the autonomy build-out — the status.html gap backlog (#102–#109 / PW-G001–G018, minus the parked apply-path G014) is unparked; codex builds + self-sequences. Engineering autonomy track only — the strategy/discovery track (V1/V4, install-count) is unchanged.
Delegated controller-merge — a controller identity may merge on a green, head-bound Patchwarden contract + handoff, within repo mission/vision scope. (Operator delegates their merge authority to the controller — not Patchwarden merging itself.)

The one judgment call I made (override at merge if you disagree)

I scoped "mission/vision" = ordinary product delivery, and kept the hard-manual classes human (secrets / workflow / runtime / auth / branch-protection / Patchwarden policy+automerge machinery) — because those are governance machinery, not "delivery," and you've always held them sacred. Other invariants kept: D20 intact (Patchwarden still never merges itself), no self-approval (merging identity ≠ PR author), fail-closed on stale/missing evidence.

If you actually want auto-merge to reach those hard-manual classes too, say so and I'll widen D24 — but I'd flag that as the genuinely irreversible move.

Safety

Docs-only; 326 tests green. Supersedes the "M2-gated backlog" tags on #103–#109; amends D21; extends D20's operator-merge into delegated controller-merge within the stated bounds.
You can narrow/revoke anytime; trust tiers (#108) + incident log (#95) provide the feedback to tighten.

Operator merge only — merging this PR is your ratification of the posture.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

## What Records your 2026-06-23 A/B/C decisions as **D24**, so the bold autonomy push has authoritative, auditable bounds (and codex knows exactly how far "go big" goes). ## The decision (two parts) 1. **M2 gate lifted for the autonomy build-out** — the status.html gap backlog (`#102`–`#109` / PW-G001–G018, minus the parked apply-path G014) is **unparked**; codex builds + self-sequences. *Engineering autonomy track only — the strategy/discovery track (V1/V4, install-count) is unchanged.* 2. **Delegated controller-merge** — a controller identity may merge on a green, head-bound Patchwarden contract + handoff, **within repo mission/vision scope**. (Operator delegates *their* merge authority to the controller — not Patchwarden merging itself.) ## The one judgment call I made (override at merge if you disagree) I scoped **"mission/vision" = ordinary product delivery**, and kept the **hard-manual classes human** (secrets / workflow / runtime / auth / branch-protection / Patchwarden policy+automerge machinery) — because those are governance machinery, not "delivery," and you've always held them sacred. Other invariants kept: **D20 intact** (Patchwarden still never merges itself), **no self-approval** (merging identity ≠ PR author), **fail-closed** on stale/missing evidence. If you actually want auto-merge to reach those hard-manual classes too, say so and I'll widen D24 — but I'd flag that as the genuinely irreversible move. ## Safety - Docs-only; **326 tests green**. Supersedes the "M2-gated backlog" tags on #103–#109; amends D21; extends D20's operator-merge into delegated controller-merge within the stated bounds. - You can narrow/revoke anytime; trust tiers (#108) + incident log (#95) provide the feedback to tighten. **Operator merge only** — merging this PR is your ratification of the posture. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Rows
Columns